Diversity Measures: Domain-Independent Proxies for Failure in Language Model Queries

TL;DR

This paper introduces domain-independent diversity measures based on entropy, Gini impurity, and centroid distance to predict errors in large language model responses, applicable across various tasks and settings.

Contribution

The paper proposes novel diversity-based error prediction measures that do not rely on domain-specific information, validated through extensive experiments across datasets and prompting methods.

Findings

Diversity measures strongly correlate with model failure probability

Measures are effective across multiple datasets and temperature settings

Applicable to few-shot prompting, chain-of-thought reasoning, and error detection

Abstract

Error prediction in large language models often relies on domain-specific information. In this paper, we present measures for quantification of error in the response of a large language model based on the diversity of responses to a given prompt - hence independent of the underlying application. We describe how three such measures - based on entropy, Gini impurity, and centroid distance - can be employed. We perform a suite of experiments on multiple datasets and temperature settings to demonstrate that these measures strongly correlate with the probability of failure. Additionally, we present empirical results demonstrating how these measures can be applied to few-shot prompting, chain-of-thought reasoning, and error detection.